Demo: Zava Media AI Assistant
Multi-Agent Architecture
for Image & Video Processing - Overview

Last updated: 2026-01-23

List of References (Click to expand)

• Foundry Models sold directly by Azure - models available
• Timelines for Foundry Models - retirement dates
• Azure OpenAI in Microsoft Foundry model deprecations and retirements - deprecation Date
• Use model router for Microsoft Foundry - model-router LLMs
• Model summary table and region availability - table summary
• Baseline architecture for an Azure Kubernetes Service (AKS) cluster
• Run your functions from a package file in Azure
• What is Microsoft Translator Pro?
• Model leaderboards in Microsoft Foundry portal (preview)
• AI Leaderboards - general ref
• How to Stream Agent Responses
• How to enable Live Streaming over Direct Line for a Copilot Studio - deployed agent?
• Azure OpenAI Responses API
• Foundry Control Plane: Managing AI agents at scale | BRK202

!!! info "Demo scope" This repository contains a demo of Zava Media AI Assistant, a hybrid system using 2 Azure AI Agents (via Azure AI Agents Service) for conversational orchestration and cropping, with code-based orchestration for other media tasks (video, image generation, document processing). It features a fully automated "Zero-Touch" deployment pipeline orchestrated by Terraform, which provisions infrastructure, creates specialized AI agents in MSFT Foundry, and deploys the complete application stack. Feel free to modify this as needed; it is a reference. Please refer to TechWorkshop L300: AI Apps and Agents, and contact Microsoft Sales and Support for additional guidance.

E.g

!!! info "Deployment duration" The deployment process typically takes 15-20 minutes.

1. Adjust terraform.tfvars values.
2. Initialize Terraform with terraform init. Learn more about the deployment process.
3. Run terraform apply. This automatically handles all deployment, including agent creation and configuration.

Deployment and reference guides

Infrastructure

Terraform deployment template

Review the template files, required values, and deployment commands.

Operations

Troubleshooting guide

Resolve setup, authentication, permissions, provider, and state issues.

Key Features

!!! warning "Regional deployment"

• Multi-Region Deployment: Sweden Central hosts 4 models + 2 agents, and East US hosts 1 model.
• All models use the GlobalStandard SKU for optimal performance and availability.

For example East US & Sweden Central:

East US	Sweden Central

• Hybrid Agent Architecture: 2 Azure AI Agents for chat-based orchestration + code-based orchestration for media processing
• Multi-Region Deployment:
  • Sweden Central: 4 models + 2 agents
    • Models: model-router, GPT-4o, Sora, FLUX.1-Kontext-pro
    • Agents: zava-media-orchestrator, vision-analyst
  • East US: 1 model (no agents)
    • Models: FLUX.2-pro
• 2 Azure AI Agents (chat-based via Responses API):
  • zava-media-orchestrator: Central request router using model-router chat model. Routes to 18+ other models
  • vision-analyst: Object detection and coordinate analysis using GPT-4o chat model with vision (provides JSON coordinates via HTTPS). ~ Analyzes images to detect objects and return bounding box coordinates as JSON. Application code handles actual image manipulation (cropping, resizing, etc.) using the provided coordinates.
• Code-Based Orchestration for generation tasks:
  • Video Generation: Direct calls to Sora (Sweden Central). ~ Video generation model (not used by agents, called directly via code)
  • Image Generation: Direct calls to FLUX.1-Kontext-pro (Sweden Central) and FLUX.2-pro (East US) ~ Image generation model (not used by agents, called directly via code).
• OSS Baseline (Open-Source): Includes an in-app OSS baseline with optional Diffusers worker on Azure (AKS or oss_azure_worker_url_override) for more realistic output with open source libraries.
• Real-Time Image Processing: Upload or paste images directly into the chat for immediate agent action
• Real MSFT Foundry Agents: Integrates with MSFT Foundry to create and host persistent agents across multiple projects
• Zero-Touch Deployment: A single terraform apply command handles the entire lifecycle
• Advanced Task Coordination: Inter-agent task delegation (e.g., "Crop this, then change background, then add text")
• Dynamic Configuration: All settings managed via terraform.tfvars - no code changes needed, just add your values here

Architecture Overview

!!! info "Agent model boundary" Agents use chat models only, not image-generation models. GPT-4o is a chat model with vision: it can analyze images in a conversation but does not generate images.

How it works

1. Orchestrator Agent (model-router - chat model) receives user requests and routes appropriately.
2. Vision Analyst Agent (GPT-4o - chat model with vision) can analyze images in chat and provide object-detection coordinates as JSON.
3. Code orchestration calls generation models directly:
   • Video generation (Sora - not an agent, direct API call).
   • Image generation (FLUX.1-Kontext-pro - not an agent, direct API call).
4. Key distinction:
   • Agents = chat models (model-router, GPT-4o) for conversation and analysis.
   • Code = generation models (Sora, FLUX) for creating videos and images.
   • GPT-4o is a chat model that can analyze images, not an image-generation model.

!!! warning "Azure quota and model availability" The deployed models (model-router, GPT-4o, FLUX.2-pro, FLUX.1-Kontext-pro, and Sora) require GPU capacity and are subject to Azure quotas. If you encounter an Insufficient Quota error, request a quota increase through Azure Support.

Architecture

graph TD
    User[User] <--> UI[Media Studio UI]
    UI <--> App[FastAPI Application]
    
    App <--> Orchestrator[zava-media-orchestrator<br/>Model Router Chat Model<br/>Sweden Central]
    App <--> Vision[vision-analyst<br/>GPT-4o Chat + Vision<br/>Object Detection & Coordinates<br/>Sweden Central]
    
    App <--> CodeOrch[Code-Based Orchestration]
    
    CodeOrch --> Sora[Sora<br/>Video Generation<br/>Sweden Central]
    CodeOrch --> FLUX1[FLUX.1-Kontext-pro<br/>Image Generation<br/>Sweden Central]
    CodeOrch --> FLUX2[FLUX.2-pro<br/>Image Generation<br/>East US]
    
    subgraph "Azure AI Agents - Chat Models Only"
        Orchestrator
        Vision
    end
    
    subgraph "Sweden Central - Generation Models"
        Sora
        FLUX1
    end
    
    subgraph "East US - Generation Models"
        FLUX2
    end

Loading

Architecture Distribution:

• 2 Azure AI Agents (Sweden Central): zava-media-orchestrator (model-router), vision-analyst (GPT-4o)
• Generation Models: Sora, FLUX.1-Kontext-pro (Sweden Central), FLUX.2-pro (East US)
• Key: As now, Agents use chat models per Azure AI Agents SDK design

What Happens Under the Hood?

When you run terraform apply, the following automated sequence occurs:

1. Infrastructure Provisioning:

   • Creates Resource Group, 2 Azure AI Foundry projects (Sweden Central + East US), Key Vault, Storage Account, and Container Registry (ACR)
   • Multi-Region Model Deployment:
     • Sweden Central (4 models):
       • Model Router (Orchestrator - automatic model selection from 18+ options)
       • GPT-4o (Vision and cropping tasks)
       • Sora (Native video generation)
       • FLUX.1-Kontext-pro (Document processing and contextual understanding)
     • East US (1 model):
       • FLUX.2-pro (Background generation, thumbnail creation, artistic image manipulation)
   • All models use GlobalStandard SKU for optimal performance
   • All resources use Managed Identity for secure authentication (no API keys stored)

2. Automated Agent Creation:

   • Fully automated by Terraform: No manual intervention required
   • Installs the azure-ai-projects SDK and connects to MSFT Foundry projects in both regions
   • Creates specialized media processing agents:
     • Sweden Central: zava-media-orchestrator, vision-analyst
     • East US: No agents (Models accessed directly via code)
   • Automatically stores agent IDs in Azure Key Vault for secure access with region prefixes
   • Web app retrieves agent configuration from Key Vault automatically
   • Zero manual configuration - Terraform handles all multi-region agent deployment and setup

3. Application Deployment:

   • Builds the Docker container in the cloud (ACR Build)
   • Configures the Azure Web App with the generated Agent IDs and Managed Identity
   • Deploys the container and restarts the app

Verification

After deployment completes, verify the system:

1. Check the Web App:

   • The Terraform output will provide the application_url

   • Visit https://<your-app-name>.azurewebsites.net

   • You should see the Zava Media AI interface

     Deployed assistant walkthrough.

2. Verify Agent Architecture:

   • Go to the MSFT Foundry Portal
   • Check Sweden Central Project -> Build -> Agents:
     • Should see: zava-media-orchestrator and vision-analyst
   • Check East US Project:
     • Note: No agents are created in East US. The FLUX.2-pro model is accessed directly via code.
   • Agent IDs are automatically stored in Azure Key Vault with region prefixes and retrieved by the web app

3. Test Processing: For example:

   • Chat: Ask for information "What is GitHub Copilot?"

   • Image Upload: Upload an image and ask "Crop the main subject"

   • Background: "Change the background to a beach scene" (routed to East US for fast generation)

   • Thumbnail: "Create a thumbnail with the text 'AMAZING'" (routed to East US)

   • Multi-Step: "Crop the car, put it on a race track background, and add the text 'SPEED' in red"

   • Video:

     Video of a bottle in different environments:

     Bottle-environments video.

     "Generate a video of a Scottish terrier" (Sweden Central - Sora)

     

     Sora video-generation result.

   • Document: "Extract all text from this PDF" or "Summarize this document" (Sweden Central - FLUX.1-Kontext-pro)